Beginner's
Guide to
MCP
A Must Read
The Challenge with Context in LLMs
LLMs like Claude 4 can process massive context
windows. This allows them to understand long
documents or codebases in a single pass.
However, while having such a large context window is
impressive, the real challenge lies in effectively managing
and providing relevant context to these models in real-
time.
The Solution: MCP
MCP simplifies how LLMs connect to external data,
systems, and tools. Think of MCP as the USB-C for AI, a
universal connector that allows LLMs to seamlessly
access and integrate context without cumbersome and
redundant custom code.
How MCP Works?
MCP divides external context into three main
components:
Resources: These are read-only data sources (e.g.,
documents, API responses) that provide context to
the LLM when needed.
Tools: These are functions or actions (e.g., sending
an email, running a computation) that the model can
invoke during its operation.
Prompts: Predefined instructions or templates that
guide the LLM's behavior, providing consistent
context for tasks like Q&A or summarization.
MCP vs. Traditional Approaches
Before MCP, LLMs were integrated with external
systems through custom, often inefficient methods:
Manual Prompt Engineering: Lacking standardization,
this required creating large, complex prompts for each
new data source.
Retrieval-Augmented Generation (RAG): A system
might use a vector search to pull in relevant context and
append it to a prompt, but it lacked flexibility and was
resource-intensive.
MCP Architecture: How Data Flows?
Host: The
LLM
application
Host: The LLM application
(e.g., Claude’s desktop
app).
MCP Client: A library
within the host that
connects to external data
sources.
MCP Server: The external
system (database, API, or
tool) providing resources,
tools, and prompts in MCP
format.
Initialization: The client connects to the server, verifying
compatibility.
Capability Discovery: The client checks what tools,
resources, and prompts the server can offer.
Context Provision: The model requests and retrieves
necessary context or tools based on the user’s query.
Action and Integration: The model executes tasks (e.g.,
querying an API) and integrates the results into its
response.
